Last week, I had the opportunity to turn a list of over one hundred manual processes shared by our department leaders into a cohesive 18-month automation roadmap.
Normally, a strategic project of this scale would take weeks of analysis and planning. I had a personal deadline that made that impossible—I was heading out on vacation and had a single two-hour window to produce a credible first draft.
So, I decided to see if a parallelized, hybrid-AI workflow could accomplish the task.
In just under two hours, it worked.
The output wasn’t perfect, but it was an incredible accelerator.
This post is a breakdown of the workflow, which explicitly addresses many of the common pitfalls and debates in the agentic development community, such as context management and the “black box” problem.
The Workflow at a Glance
This isn’t a single black box; it’s a transparent, auditable assembly line with clear checkpoints for human review.
graph TD
subgraph "Step 1: Data Preparation"
A[📄 Raw Spreadsheet] --> B(Python Script);
B --> C[Clean Markdown Files];
end
subgraph "Step 2: Command & Prompt Engineering (Claude)"
D{👨💻 Human-in-the-Loop};
E[Agent Command];
D -- Run & Review --> E;
E -- Tweak Prompt --> D;
end
subgraph "Step 3: Parallel Execution (Claude)"
F[⚡️ Scaled Analysis];
G[Agent 1: Batch A];
H[Agent 2: Batch B];
I[Agent N: ...];
F --> G & H & I;
end
subgraph "Step 4: Strategic Synthesis (Gemini)"
J[Enriched Analyses];
K{🧠 Gemini 2.5 Pro};
L[Draft Roadmap];
J --> K --> L;
end
subgraph "Step 5: Finalization"
N[Final Agent: Format Timeline];
M[🗺️ Final 18-Month Roadmap];
N --> M;
end
%% Connect the Steps
C --> D;
E --> F;
I --> J;
L --> N;The Process in Detail
My approach followed five distinct steps, leveraging different tools for what they do best.
Step 1: Data Preparation – From Spreadsheet to Markdown
The initial data was a spreadsheet with over 100 processes. My goal was to break this down so an agent could analyze each one individually. I’ve learned that using an LLM for pure, structured data transformation is often less reliable than a simple script.
I used Claude Code to create a Python script to parse the sheets into clean Markdown files, creating a reliable, auditable data foundation.
graph TD
A[Raw Spreadsheet Data] --> B{Python Script};
B --> C[Individual Markdown Files <br> 1 per process];
C --> D[Spot Check / Validation];
D --> E[Clean Data Ready for Agents];Step 2: Command Development – Perfecting the “Unit of Work”
With the data prepared, I built a subagent in Claude Code to analyze a single process.
The key learning from past work is to define agents around specific tasks, not just roles. My automation-architect agent was designed not just to “act like an architect,” but to execute a specific task: evaluate a process through a company-specific lens (eg: knowing our team skillset and what tech we are using) and produce a structured one-pager analysis.
My iterative process was to:
- Draft the command with essential business context (e.g., “We use N8N,” “We can build custom agents”).
- Run and Review the output against my own expertise. A human’s value is knowing “what good looks like.”
- Iterate on the prompt until the agent’s one-pager consistently met the standard I would expect from a human analyst.
The Agent Behind the Command: automation-architect
Here are the core instructions for the specialist agent that formed the heart of my command:
automation-architect markdown file
---
name: automation-architect
description: Evaluates automation opportunities, designs workflow solutions, and makes pragmatic build-vs-buy decisions for business processes.
model: opus
color: green
---
You are a pragmatic Product Technologist and Automation Architect. You excel at evaluating automation opportunities by focusing on clear ROI, technical feasibility, and speed-to-value.
Your primary goal is to identify high-cost manual processes and design the most efficient automated solution, whether that's a low-code workflow, an existing SaaS tool, or a custom-built agentic system.
## Core Expertise
You possess comprehensive knowledge of:
- Low-code/no-code platforms (e.g., N8n, Zapier, Make)
- Pragmatic agentic workflow design (MVP-first, not enterprise-grade perfection)
- Simple API integrations and webhook architectures
- Process automation ROI calculation for startups (weeks to payback, not years)
- Rapid custom automation development (Python/Node.js scripts, serverless)
## Decision Framework
When evaluating an automation opportunity, you systematically analyze three key areas:
1. **Business Value Assessment**
- **Time & Cost Savings**: How many hours per month does this save? What is the opportunity cost of that time?
- **Error Reduction**: What is the business cost of manual errors?
- **Scalability**: Does this automation unblock future growth?
- Calculate a clear ROI including implementation and maintenance costs.
2. **Technical Feasibility Analysis**
- Map the current process flow and data dependencies.
- Identify integration points and API availability.
- Assess error handling and recovery needs.
3. **Build vs. Buy Decision Matrix**
- **Buy (SaaS)**: When an off-the-shelf solution meets >80% of core needs.
- **Low-Code (e.g., N8n)**: For medium complexity tasks with standard system integrations.
- **Custom Build (Pragmatic)**: For unique processes that provide a competitive advantage, scoped to a 2-4 week MVP.
- **Hybrid Approach**: Use a low-code platform for orchestration and custom code for the complex business logic.
## Research Methodology
Upon receiving a process to analyze, you launch parallel web searches to gather intelligence.
**Wave 1 - Solution Discovery (Simultaneous)**:
1. **Existing Products**: `"[process name]" automation software SaaS`
2. **Low-Code Workflows**: `site:n8n.io/workflows "[process keywords]"`
3. **Open Source Solutions**: `github "[process]" automation workflow opensource`
**Wave 2 - Deep Dive (Based on Wave 1)**:
- Specific vendor comparisons: `"[Product A]" vs "[Product B]" review`
- Implementation guides: `"[chosen solution]" tutorial`
- Cost analysis: `"[solution]" pricing`
## Pragmatic Evaluation Criteria
You always consider:
- **Time to Value**: Can we see a tangible benefit in under a month?
- **Maintenance Burden**: Who will maintain this and at what cost?
- **Skill Requirements**: Do we have the expertise in-house?
- **Vendor Lock-in**: How difficult would it be to migrate away from this solution?
- **Compliance & Security**: Does this meet our requirements?
## Output Structure
For each analysis, you will produce a structured markdown document that includes:
- **Executive Summary**: Your final recommendation and the estimated ROI.
- **Current State Analysis**: A breakdownStep 3: Scaling Up – From One to One Hundred in Parallel
Once I had a robust command, it was time to scale. I opened multiple Claude Code terminals and began running the command in parallel on different files.
Here, I made a pragmatic trade-off: I sacrificed some quality for speed. Instead of running each of the 100+ processes in perfect isolation, I started feeding the command entire directories grouped by department to finish within my two-hour window.
graph TD
A[Finalized Command <br> using automation-architect] --> B[Input: 100+ Process Markdown Files];
B --> C{Parallel Execution <br> Multiple Claude Terminals};
C --> D[Batch 1 Processes];
C --> E[Batch 2 Processes];
C --> F[...up to Batch N Processes];
G & H & I --> J[Collection of Detailed One-Pagers];Step 4: The Hybrid Approach – Strategic Synthesis with Gemini
After the parallel execution in Claude Code, I was left with dozens of high-quality, individually analyzed “one-pager” files. For the final, high-level strategic synthesis, I switched tools to Gemini 2.5 Pro for its ability to reason over large contexts and draw complex conclusions.
This is the key to the whole workflow. We didn’t just ask Gemini to analyze the raw data. Instead, we fed it a dataset that had been massively enriched by the prior agentic work—a collection of dossiers each containing a feasibility assessment, ROI analysis, and build vs. buy recommendation.
The prompt for Gemini was not “summarize this,” but: “You are a strategist. Given these fully analyzed initiatives, draft a cohesive 18-month roadmap, sequencing these projects logically.”
sample gemini synthesis prompt
You are a business strategist tasked with analyzing {analyses_count} detailed automation proposals to create a high-level implementation roadmap.
Our primary goal is to significantly improve operational efficiency and support company growth targets through strategic automation, focusing on maximizing the productivity of our existing teams
PLANNING CONTEXT:
We operate in standard 6-week Delivery Cycles.
The upcoming cycles are: Q4-Cycle1, Q4-Cycle2, Q1-Cycle1, Q1-Cycle2, and so on for the next 18 months.
RESOURCE CONSTRAINTS:
Automation Team (Low-Code): Assume a fixed capacity of specialists available per cycle.
Engineering Team (Custom Build): Assume a variable allocation; you must specify how many engineers are needed for each project.
Plan for realistic capacity and concurrent work within each 6-week cycle.
PRIORITIZATION FRAMEWORK:
High-Impact Initiatives FIRST: Prioritize automations that directly enable core business functions (e.g., sales, product delivery).
Productivity Multipliers SECOND: Prioritize work that frees up significant time for high-value activities (e.g., removing administrative burdens).
Foundation & Optimization LAST: Foundational tech debt or minor optimizations should be scheduled after the highest-impact work.
Your task is to find the optimal sequence of initiatives that delivers the most business value within these constraints.
Write a professional EXECUTIVE MEMORANDUM ready for leadership review:
MEMORANDUM
TO: Executive Leadership
FROM: [Your Name/Team]
DATE: [Current Date]
RE: Proposed Automation Roadmap for Q4 2025 - Q3 2026
1. Recommendation
[State the specific recommendation in 1-2 sentences: implement X automations in Y order over the next Z cycles to achieve a specific, high-level business outcome.]
2. Strategic Rationale
Business Impact: [2-3 sentences explaining how this sequence directly solves key business problems and removes critical bottlenecks.]
Productivity Multiplication: [2-3 sentences on how this roadmap enables existing staff to handle significantly more volume or focus on higher-value work.]
Optimal Sequencing: [2-3 sentences on why this specific order maximizes impact, creates compound benefits, or addresses foundational issues first.]
3. Implementation Timeline
Q4 2025 (Cycle 1 & 2)
Focus: [e.g., Foundational Wins]
Resource Allocation:
Automation Team: [X] specialists
Engineering Team: [Y] engineers needed
Deliverables:
Cycle 1: [Project A], [Project B]
Cycle 2: [Project C]
Q1 2026 (Cycle 3 & 4)
Focus: [e.g., Scaling Operations]
Resource Allocation:
...Deliverables:
...
[Continue for the full 18-month roadmap]
4. Financial Impact Summary
Metric Current State Projected 18-Month Impact
Operational Efficiency Baseline +X%
Processing Speed Y hours/days -Z%
Cost Savings (Annualized) $0 ~$[Value]/year
Export to Sheets
Investment Required: [Total] developer-weeks (using existing teams)
Estimated Payback Period: [X] months
5. Risk Assessment
Primary Risk: [Identify the main risk to the plan, e.g., "Dependency on a single API's stability."]
Mitigation: [Specific mitigation plan, e.g., "Develop an abstraction layer and robust error handling."]
6. Next Steps
Upon approval, we will proceed with finalizing resource allocation for Q4 and begin implementation.
CRITICAL INSTRUCTIONS:
THINK IN DELIVERY CYCLES: All timelines must map to the 6-week cycles. Be explicit about team allocation.
SYNTHESIZE FOR IMPACT: Your primary goal is to find the sequence that delivers the most significant business value the fastest
BE SPECIFIC AND ACTIONABLE: Name the exact processes from the analyses provided.graph TD
A[Collection of Detailed One-Pagers <br> from Claude] --> B{Input to Gemini 2.5 Pro};
B --> C["Prompt: Act as a strategist, <br> sequence these initiatives, <br> create a roadmap."];
C --> D[Draft Strategic Roadmap];Step 5: Finalizing the Timeline
The draft roadmap from Gemini provided the strategic sequencing. The final step was to map this strategy onto our company’s specific operating cadence. I gave this draft to a final agent and instructed it to “Map these projects onto our six-week cycle timeline,” which produced the final, shareable document.
The Result and The Reality Check
The roadmap was a powerful starting point. It had flaws, of course—the agents made assumptions about team sizes and hours spent, which skewed some ROI calculations.
But this doesn’t concern me because the system is now a turnkey, repeatable process. We can easily add context to prevent those assumptions and re-generate the roadmap.
The dollars in token costs for this 2-hour workflow are negligible compared to the fully-loaded salary cost of a senior strategist working for weeks to produce a similar first draft. The value is in the massive compression of time and effort.
Learnings and The Future
This exercise reinforced that the most significant flaw—making incorrect assumptions—is also the most fixable. The next version of this system would be updated to ask clarifying questions instead of guessing, making the human an active partner in the analysis.
Ultimately, this was about applying an engineering workflow—parallelism, automation, iterative refinement—to a product and strategy problem. This approach is exceptionally powerful for strategic and greenfield tasks, where problems can be neatly compartmentalized. It proves that the impact of these tools goes far beyond the command line.
Discover more from zach wills
Subscribe to get the latest posts sent to your email.